Decoding multimodal social media data with LLM to reveal agricultural landscape recreational services in Chengdu, China.

Agricultural landscapes in metropolitan regions are increasingly expected to provide recreational opportunities alongside productive and ecological functions. However, existing studies have rarely integrated public perception with actual recreational use, and the nonlinear relationships among landscape structure, accessibility, and recreational services remain insufficiently understood. Using Chengdu as a case study, this study integrated social media text-image content and behavioral trajectory data to construct an Agricultural Landscape Recreational Services (ALRS) assessment framework. A Geographically Weighted Random Forest (GWRF) model combined with SHAP interpretation was further used to identify key environmental drivers and nonlinear relationships. The results showed that ALRS was influenced by landscape structure, accessibility, and topographic conditions. PD, accessibility to agricultural recreational facilities, Elevation, and accessibility to tourist attractions were identified as the most influential variables. Environmental variables exhibited pronounced nonlinear and threshold effects, particularly in relation to landscape structure and accessibility. K-means clustering further identified four ALRS zones with distinct spatial characteristics and development constraints. Overall, The findings suggest that ALRS in metropolitan agricultural landscapes was shaped by context-dependent interactions among landscape configuration, topographic conditions, accessibility, and recreational resources, rather than by a uniformly synergistic relationship among individual drivers. This study provides methodological support for ALRS assessment and differentiated spatial optimization in metropolitan regions.